A re-evaluation of the term spread as a leading indicator

Forecasting the evolution path of macroeconomic variables has always been of keen interest to policy makers and market participants. A common tool used in the relevant forecasting literature is the term spread of Treasury bond yields. In this paper, we decompose the term spread into an expectation and a term premium component and evaluate the informational content of each component in forecasting the GDP growth rate and inflation in various forecasting horizons. In doing so, we employ alternative decomposition procedures and introduce the Support Vector Regression (SVR) methodology from the field of Machine Learning, coupled with linear and non-linear kernels as a novel forecasting method in the field. Using rolling windows in producing point and conditional probability distribution forecasts we find that neither the term spread, nor its decomposition components possess the ability to accurately forecast output growth or inflation. Our findings extend the existing literature, since they are focused on an explicit out-of-sample evaluation in contrast to most existing empirical studies that produce only in-sample forecasts. To strengthen our findings, we also consider several control variables suggested in the relevant literature without significant qualitative differences from the initial results. The main innovation of our approach stems from the use of the non-linear Support Vectors Machine methodology, that is introduced for the first time in this line of research for forecasting out-of-sample.